Glossary

Plain-English definitions of the AI terms you will see on this site. Each one is under forty words so it is quick to read between patients.

Agent: An AI tool that can take a series of steps on your behalf — for example, looking something up and then writing a summary. Slower, but useful for multi-step jobs.
AI (Artificial Intelligence): Computer software that can perform tasks people usually associate with thinking, such as writing summaries, answering questions, or recognising patterns.
API: A way for one piece of software to talk to another. You will see this term in tools that connect AI to a hospital system. As a user, you can mostly ignore it.
Chatbot: A program you can hold a written conversation with. You type a question or instruction, and it writes a reply.
ChatGPT, Claude, Gemini: Three of the best-known AI chat tools, made by OpenAI, Anthropic, and Google respectively. They are broadly similar in what they can do. Pick whichever is approved at your organisation.
Context window: How much text the AI tool can keep in mind during a conversation. If a chat goes on long enough, earlier messages may be forgotten. Start a fresh chat for a fresh topic.
De-identified data: Clinical information with all names, dates of birth, identifiers, and other patient details removed. Use de-identified data — or made-up data — when trying an AI tool.
Embedding: A numeric fingerprint of a piece of text, used by AI tools to find similar passages. A technical detail you will rarely need to think about as a clinician.
Enterprise version: A paid version of an AI tool sold to organisations. It usually promises that your inputs will not be used to train future models. Check the contract before assuming.
Fine-tuning: Re-training an AI model on a specific set of examples to make it better at one job. Done by software teams, not users. You will not need to do this.
Generative AI: AI that produces new content — text, images, audio — rather than only classifying or scoring existing content. The chat tools in these lessons are all generative AI.
Guardrails: Built-in rules that stop an AI tool from producing certain content, for example dangerous medical advice. Useful, but never assume guardrails are perfect.
Hallucination: When an AI tool produces an answer that sounds confident but is factually wrong. Always double-check clinical facts against a trusted source.
Inference: The technical name for the moment an AI tool generates a reply to your prompt. You may see it in pricing pages. Practically, it just means "running the model".
Jailbreak: A prompt designed to trick an AI tool into ignoring its safety rules. Worth knowing the term exists; not something a clinician should attempt.
Large language model: The kind of AI that powers tools like ChatGPT. It is trained on large amounts of written text and predicts likely sentences in response to your prompt.
Model: The underlying AI engine inside a tool. Different models have different strengths; newer versions tend to be more accurate.
Multimodal: A tool that can read more than one kind of input — text plus images, for example. Useful for asking questions about a photograph of a chart or document.
Open-source model: A model whose internals are published, so it can be run on your own computers. Relevant if your organisation wants to keep data fully in-house.
PHI (Protected Health Information): Any information that could identify a patient and link them to their health. Names, dates of birth, hospital numbers, addresses, and photos all count. Do not paste it into general-purpose AI tools.
Privacy mode: A setting in some AI tools that prevents your conversations from being used to train future versions. Worth turning on; not a substitute for de-identifying data.
Prompt: The instruction you type into an AI tool. A clearer prompt usually gives a better answer.
Prompt engineering: The skill of writing prompts that produce reliably good answers. Mostly a matter of being specific about audience, length, tone, and structure.
RAG (Retrieval-Augmented Generation): A technique where an AI tool looks something up — for example in a guideline document — before answering. Used in some clinical tools to ground replies in trusted sources.
Reasoning model: A newer kind of AI tool that "thinks" for longer before answering. Slower but better at multi-step problems such as working through a differential.
Safety filter: Software that checks AI output for harmful content before showing it to you. Helpful, but designed for general harms, not clinical safety. The reader is still the safety check.
System prompt: A hidden instruction that shapes how an AI tool behaves throughout a conversation — for example, "always answer as a senior clinician would". You can sometimes set this yourself.
Temperature: A setting that controls how predictable or creative an AI tool's answers are. Lower temperature is steadier and more factual; higher is more varied. Most tools default to a sensible middle.
Token: A small chunk of text (roughly four characters) that the AI reads or produces. Tools often have limits measured in tokens.
Training data: The text or images an AI tool learned from. It shapes what the tool knows and where it is weakest.
Transcription model: An AI tool that converts spoken words into written text. Useful for dictating clinic notes; the same patient-privacy rules apply.
Vision model: An AI tool that can look at an image — a chart, an x-ray, a photograph — and describe or analyse it. Treat clinical outputs with the same care as text outputs.